feat(ui): announce blocking actions in the header and add a timing page with Buckaroo's spans - #318

…eue and for trace labels

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

…a timing page

activity.ts holds the pure state: a reducer for action_start and action_end, the now and last labels, the wait a load_expr POST spent in Buckaroo's queue, and the timeline layout. SSEContext keeps it without bumping version; ActivityStrip shows it in the header; TimingPage lays the companion's actions and Buckaroo's spans on one axis.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

…burst of activity

The rows lacked the flex layout the shared waterfall classes give, and one axis for the whole window made a half-second action a sliver next to another half a minute later. Rows within 2 s of the latest end share an axis. The note on a slow load_expr now says when Buckaroo began handling it, since the gap also holds the request's transfer and parsing, not only queueing.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

…em over SSE, and serve /api/timing

tallyman_core/activity.py tracks running and finished actions and tells listeners at each start and end. load_session announces preparing the entry and the load_expr POST; the diff, promote, save-code, reset, staleness and recalc routes announce themselves. The companion forwards the events over SSE and GET /{project}/api/timing returns the actions beside the spans Buckaroo posted back.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

…trip and timing page

Python: a malformed telemetry span 500s /api/timing; reading spans parses the whole file; refused requests (read-only,
unknown project) are announced as failed actions; a forced reload, the klass reload and the diff view's load_expr are
not announced; the async routes hold the event loop through their blocking work, so action_start is held until the
work is done; the hello event carries no activity snapshot.

Vitest: formatDuration prints the wrong unit at its edges; a zero-length row sticks out of its track; seeding the
header from a snapshot, the project filter, span keys and alias labels do not exist yet.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

…d fix the timing page's labels

The header's state is replaced by the activity snapshot on each SSE hello (first connect and every reconnect), so an
action_end missed while disconnected no longer leaves a 'now:' running for good, and /api/timing is no longer read just
to seed it. Activity is reset when the project changes and live events are filtered to the open project. Activity moves
into its own context so an action event re-renders the strip and the timing page, not every useSSE() consumer.

formatDuration picks its unit from the rounded value. layoutTimeline keeps a minimum-width row at the end inside the
track, and the grid-load waterfall uses it instead of its own copy of the math. Timing rows are keyed by what the span
is rather than its index, and aliases are labelled with their version and refetched when entries change.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

…pshot on hello, and read telemetry cheaply

The async routes ran their blocking steps on the event loop, so the action_start for a slow promote or save was held
until the work was done. promote_diff and put_code now run diff_keys, cached_result_expr and build_and_persist on the
thread pool, and reset, recalc, auto-recalc and /internal/notify do the same for reload_project_sessions.

The SSE hello event carries the project's running and recent actions, which the SPA replaces its header state with.
A request for an unknown project, or a write in serve mode, is refused before it is announced. A forced reload, the
klass reload and the diff view's load_expr are announced like the other POSTs that queue behind Buckaroo's stats.

/api/timing skips span records whose times are not finite numbers instead of failing, and reads telemetry.jsonl
backward only as far as its window instead of parsing the whole file on every poll.

The test that added a listener to the process-wide tracker now removes it.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>